AI agents hit 85% on computer tasks, undercutting $10 BPO labor

AI agents now complete 85 percent of standard computer tasks, surpassing the 72 percent human baseline, while running at $6 to $8 per hour — below the $10 fully loaded cost of offshore BPO labor.

"The question has shifted from 'can agents use a computer?' to 'can they reliably do this job?'" Fabrizio Serafini, early AI investment partner at a16z, said in research published this month.

The OSWorld-Verified benchmark, which tests agents across Windows, macOS, and Ubuntu desktop environments, shows the top model Claude Fable 5 scoring 85 percent versus roughly 72 percent for human testers. A year ago, the best model scored 42 percent. Serafini noted that production deployments rarely use consumer products — enterprises call raw APIs and build their own sandbox VMs, orchestration logic, and retry mechanisms.

The economics are now compelling enough to reshape the $300 billion global BPO market. Companies with heavy offshore labor exposure face margin compression, while vertical AI automation startups stand to capture value as the competitive moat shifts from model capability to application-layer workflow integration.

The Infrastructure Layer Becomes the Moat

Serafini's research, based on interviews with production teams, found that raw UI navigation is becoming a commodity. The durable advantage sits in the layers above the model: context management, permissions, process knowledge, validation, and error handling. One operator running millions of automated tasks monthly couldn't recall which model powered their system — the vendor's ability to swap models underneath without breaking the workflow was the value.

A consumer data platform runs 15 million to 20 million automated portal interactions monthly, using agents as a self-healing fallback when retailer portals change their UI. A global systems integrator deployed 27 live workflows processing 1,500 to 2,100 IT tickets daily, targeting redeployment of 20 to 25 percent of staff from low-margin managed services contracts.

The 85 percent completion rate means 15 of every 100 tasks still fail, and back-office processes require every step to succeed. Serafini documented failure scenarios: an agent misreading "net 60" as "net 30" in an ERP system produces a record that looks normal until the wrong invoice ships. An insurance claim submitted and acknowledged by the system can silently stall when a claims adjuster calls two days later for verification. "More intelligent models cannot fix a process where the truth arrives as a phone call on someone's desk a week later," he said, "unless the execution framework is designed for these edge cases from the start."

Cost Curve Crosses the Break-Even Line

The $6 to $8 per hour agent cost assumes the most expensive configuration — a frontier model running frame-by-frame screenshots. Optimized harnesses offload repeatable steps to deterministic code, cutting blended costs further. Against India offshore BPO at $10 per hour fully loaded, and US back-office labor at $30 to $45 per hour, agents offer roughly 70 to 80 percent gross margin savings versus domestic staffing. Serafini cautioned the figures are order-of-magnitude references, not precise quotes.

The shift favors vertical-specific AI startups that embed workflow knowledge — internal terminology, escalation paths, validation rules — into their products. Labs including Anthropic and OpenAI, plus infrastructure startups Mechanize, Habitat, Fleet, Chakra, Deeptune, Matrices, and Originator, have poured hundreds of millions into computer-use reinforcement learning environments over the past year. Standard Intelligence's 11-million-hour video dataset and 30-frames-per-second universal computer action model signal the direction of travel.

For investors, the re-rating risk is concentrated in BPO and IT services names with significant offshore exposure, including Genpact, Infosys, Wipro, Concentrix, and TTEC. The counter-position is in AI workflow automation and enterprise context-layer companies that own the integration layer rather than the model.

This article is for informational purposes only and does not constitute investment advice.